Initial implementation

Included:

• Create a tarball from a given "dataset" directory (such as a Mercurial repository) and upload it with rsync
• Download tarballs via HTTPS, check their SHA-256 hash, keep them in cache
• Extract tarballs into temporary directories, run a shell script, and report the time that takes

Planned:

• Compile a configurable changeset of Mercurial to run in benchmarks
• Cache installs of compiled Mercurial
• Choose a Mercurial "variant" (it's called a module policy) among py, py+c, py+c+rust, rhg, etc.
• Take benchmarking ideas from hyperfine
  • Run the same benchmarks multiple times, measure average duration and standard deviation
  • Subtract the time to spawn /bin/sh and run an empty shell script
  • Add a "warm-up" run
  • Auto-adaptive number of runs?
  • Detect statistical outliers?
• Save results to files, together with environment info
• Compare pairs of result sets (report for each dataset ±X% faster and ±N milliseconds faster)